Hi @Nick_Hughes ,

I am very late to the party, but I was digging in the internet to find more people discussing a relatable problem for which I am on my way building a definitive solution, and came across your post from 3 years ago. Times have changed and you might have figured a way to solve the solutions. But for the people to prevent such issues in the future, I am also building on a solution. In 2026, most people rely on AI Agents for building such solutions. But for multi-table relationship and maintaining statistical fidelty of the data along with generation thousands of rows, the token cost might be too high. So I was working on a mathematical solution for it, along with integration with the latest workflow such as MCPs and Natural Language to Schema and then Schema to Dataset generation.

So if anyone would like to checkout what I am building and give me feedback on where the library is currently weak and where to improve, this is my library: https://github.com/rasinmuhammed/misata

Misata is built exactly for a usecase of generating fake data without having reference data, but also on maitaining referential integrity + outcome conformant synthesis so that the data could convey the story we want it to convey.